Probabilistic downscaling of GCM scenarios over southern India

  • Published on
    09-Aug-2016

  • View
    218

  • Download
    2

Embed Size (px)

Transcript

<ul><li><p>INTERNATIONAL JOURNAL OF CLIMATOLOGYInt. J. Climatol. (2012)Published online in Wiley Online Library(wileyonlinelibrary.com) DOI: 10.1002/joc.3509</p><p>Probabilistic downscaling of GCM scenarios over southernIndia</p><p>N. Vigaud,a* M. Vracb and Y. Caballeroca Service EAU, Bureau de Recherches Geologiques et Minie`res (BRGM), Montpellier, France</p><p>b Laboratoire des Sciences du Climat et de lEnvironnement (LSCE-IPSL), CEA/CNRS/UVSQ, Gif-sur-Yvette, Francec Service Geologique Regional, Bureau de Recherches Geologiques et Minie`res (BRGM), Montpellier, France</p><p>ABSTRACT: The cumulative distribution function transform (CDF-t) is used to downscale daily precipitation and surfacetemperatures from a set of Global climate model (GCM) climatic projections over southern India. To deal with the fullannual cycle, the approach has been applied by months, allowing downscaled projections for all seasons. First, CDF-t isvalidated over a historical period using observation from the Indian Meteorological Department (IMD). Resulting highresolution fields show substantial improvements compared to original GCM outputs in terms of distribution, seasonal cycleand monsoon means for arid, semi-arid and wetter regions of the subcontinent. Then, CDF-t is applied to GCM large-scalefields to project rainfall and surface temperature changes for the 21st century under the IPCC SRES A2 scenario. Theresults obtained show an increase of rainfall, mostly during the monsoon season, while winter precipitation is reduced, andsuggest a widespread warming especially in the winter and post-monsoon season. Copyright 2012 Royal MeteorologicalSociety</p><p>KEY WORDS downscaling; climatic scenarios; southern India</p><p>Received 14 June 2011; Revised 22 March 2012; Accepted 3 April 2012</p><p>1. Introduction</p><p>Associated with fast social and economic growth locally,climate changes are likely to seriously impact India. Asnoted by Kundzewicz et al. (2007), southern India isalready a water-stressed region. Climate changes havealready been observed over the subcontinent, whereincreases of 0.40.6 C have occurred over the past cen-tury together with the annual mean temperature warmingmost pronounced during post-monsoon and winter peri-ods (Rupa Kumar et al., 2006; Bhattacharya, 2007). Interms of precipitation, Cruz et al. (2007) have observedfor the last decades that extreme summer monsoon rainsincrease over northwest India and the number of rainydays decrease along the east coast. While these projec-tions are subject to large uncertainties (Paeth et al., 2008),the potential impacts on water resources in India still needto be assessed depending on their location.</p><p>Global climate models (GCMs) are nowadays the onlytool at disposal to investigate future climate variability.However, GCM projections cannot be used directly forimpact studies due to the coarse resolution of GCM out-puts which are not suited for regional assessments (Wilbyet al., 2004). Therefore, downscaling methods have beendeveloped to go from large-scale data to local-scaledata. The dynamical approach consists of using regional</p><p>* Correspondence to: N. Vigaud, Service EAU, Bureau de RecherchesGeologiques et Minie`res (BRGM), Montpellier, France.E-mail: nicolas.vigaud@gmail.com</p><p>climate models (RCMs) to resolve physical equationsof atmospheric regional dynamics (Wood et al., 2004).RCMs are, however, domain dependant and computation-ally expensive, which restricts their use for many applica-tions. The statistical approach, on the other hand, refers tostatistical relationships between large-scale GCM featuresand local-scale climatic variables (such as precipitation ortemperature, for instance). Statistical downscaling meth-ods (SDMs) are quite flexible and generally require lesscomputational costs. Such advantages make them partic-ularly attractive for regional impact studies. SDMs canbe classified into three major categories: transfer func-tions, weather typing and weather generators. Transferfunctions are based on direct quantitative relationshipsbetween predictand and predictors through regression-like methods (Prudhomme et al., 2002). Weather typingapproaches consist in the grouping (or clustering) ofatmospheric circulations in relation to local meteorolog-ical variables (Vrac et al., 2007), while weather genera-tors are stochastic models simulating local-scale variablesbased on their probability density function, whose param-eters depend on large-scale information (Hughes et al.,1999; Wilks and Wilby, 1999; Vrac and Naveau, 2007).Worthnotingly, a common assumption to all SDMs is thatthe physical relationships underlying the statistical rela-tionships identified over a historical period remain validfor the future climate scenarios to be downscaled.</p><p>Several SDMs have already been used for downscal-ing rainfall over India, such as relevance (Ghosh and</p><p>Copyright 2012 Royal Meteorological Society</p></li><li><p>N. VIGAUD et al.</p><p>Mujumdar, 2008; Mujumdar and Ghosh, 2008) and sup-port vector machine (Tripathi et al., 2006; Anandhi et al.,2008, 2009), fuzzy clustering (Ghosh and Mujumdar,2007; Mujumdar and Ghosh, 2008) and also conditionalrandom field distribution (Raje and Mujumdar, 2009). Inthese recent studies, methods are generally validated byexamining the quantiles, cumulative and probability dis-tribution functions (CDFs and PDFs) of daily rainfall aswell as dry/wet spells lengths at local scale. However,SDMs can also be used to model relationships betweenlarge-scale and local-scale statistical characteristics andcan be referred in this context as probabilistic downscal-ing methods (PDMs). The cumulative distribution func-tion transform (CDF-t) presented in Michelangeli et al.(2009) has the advantage of directly dealing with andproviding CDFs. This method is used in this paper todownscale GCM projections from the Intergovernmen-tal Panel on Climate Change Fourth Assessment Report(IPCC AR4) in order to investigate projected changesin both rainfall and surface temperatures over south-ern India. The purpose of this study is to documentthe use of CDF-t for providing regional precipitationand surface temperature changes as derived from severalmedium-term GCM projections under the Special Reporton Emission Scenarios (SRES published by the IPCC) A2scenario (horizons 20402060). Most of the recent statis-tical downscaling studies of future climate scenarios overIndia were restrained to the monsoon season (Tripathiet al., 2006; Mujumdar and Ghosh, 2008) or to specificwatersheds (Ghosh and Mujumdar, 2007; Anandhi et al.,2008, 2009). This paper aims to present results from theCDF-t probabilistic downscaling method applied not onlyto the JuneSeptember (JJAS) monsoon period alone butto the full annual cycle, and for a domain covering thewhole of southern India. In the next section, the dataused and the downscaling method are presented. Then,the CDF-t approach is validated on a historical periodin Section 3, prior to applying the method to downscalefuture scenarios from IPCC AR4 experiments in Section</p><p>4. Discussion and conclusions are then gathered at theend in Section 5.</p><p>2. Data and methods2.1. GCMs data and observationsIn order to investigate climate change impacts over south-ern India, several GCMs from the last IPCC AR4 exercisehave been used. Kripalani et al. (2007) identified a set ofseven GCMs as most reliable regarding Indian monsoonrainfall based on their representation of the mean sea-sonal cycle in phase and amplitude. Regarding the avail-ability of daily standard outputs from the Program forClimate Model Diagnosis and Intercomparison (PCMDI)database, five GCMs have been retained (see Table I).However, at the time of this study MIROCMR dailysurface temperatures for the SRES A2 scenario werenot available from the PCMDI archives. Consequently,only four GCMs will be used for the downscaling ofsurface temperatures. Except for CGCM3, which incor-porates heat and water fluxes adjustments, these new-generation GCMs do not use surface flux correction tomaintain a stable climate in their control runs. Moredetails about the model components can be found athttp://www.pcmdi.llnl.gov/ipcc/model-documentation/.</p><p>For the purpose of this study, simulated daily rainfalland surface temperatures have been considered from bothhistorical experiments (run 20cm3 for the 19711999period) and future projections under the greenhousegas emission scenario A2 (run A2 for the 20462065period). The A2 storyline, based on high population andregionally oriented economic growth with significant andwidespread decline in fertility (Nakicenovic et al., 2000),actually describes a very homogeneous world with slowereconomic and technical changes than other scenarios.</p><p>Local-scale observations from the Indian Meteorolog-ical Department (IMD) are also used in this study. IMDdaily rainfall is available from 1971 to 2005 on a half-degree grid and surface temperatures from 1969 to 2005</p><p>Table I. Climate models and their references participating in the IPCC AR4 experiments (adapted from Kripalani et al. (2007)).Abbreviated acronyms are used in the text to identify each GCM.</p><p>No. Originating group Country IPCC ID Abbreviation References</p><p>1 Canadian centre for climatemodelling</p><p>Canada CGCM3.1 (t47) CGCM3 Flato et al. (2000)</p><p>2 Meteo-France/Centre Nationalde Recherches Meteorologiques</p><p>France CNRM-CM3 CNRM3 Salas-Media et al.(2006)</p><p>3 Max Planck Institute forMeteorology</p><p>Germany ECHAM5/MPI-OM ECHAM5 Jungelaus et al.(2006)</p><p>4 Bjerknes Centre for ClimateResearch</p><p>Norway BCCR-BCM2.0 BCCR2 Furevik et al.(2003)</p><p>5 Centre for Climate SystemResearch (The University ofTokyo) National Institute forEnvironmental Studies andFrontier Research Centre forGlobal Change (JAMSTEC)</p><p>Japan MIROC3.2 (medres) MIROCMR K-1 ModelDevelopers (2004)</p><p>Copyright 2012 Royal Meteorological Society Int. J. Climatol. (2012)</p></li><li><p>PROBABILISTIC DOWNSCALING OF GCM SCENARIOS OVER SOUTHERN INDIA</p><p>at one degree resolution. It leads to sets of 729 and210 grid points for rainfall and surface temperature,respectively, over a region comprised between 7.5 N and20.5 N and 70 E and 83 E to which we will refer in thefollowing as southern India. More details about the IMDhigh resolution daily gridded rainfall and surface temper-atures dataset can be found in Rajeevan and Bhate (2008)and Srivastava et al. (2008), respectively.</p><p>2.2. CDF-t methodThe downscaling approach chosen here is the CDF-t (Michelangeli et al. (2009)) which can be seen asan extension of the quantile-matching method. CDF-toffers the advantage to directly deal with and provideCDFs. In its non-parametric form it does not make anyassumption on the shape or family of distribution andthus can be applied separately to both rainfall and surfacetemperatures in the context of this study. CDF-t hasalready been successfully used to downscale GCMs andreanalyse 10 m wind over France by Michelangeli et al.(2009).</p><p>Let Fh stand for the CDF of observed local dataat a given weather station (or IMD grid cell) over ahistorical time period h, Gh the CDF of GCM outputs bi-linearly interpolated at the station location for the sameperiod, Ff and Gf their equivalent for the future periodconsidered. The method is based on the assumption thatthere exists a transformation T translating the CDF ofa GCM variable (predictor) into the CDF representingthe local-scale climate variable (predictand) at the givenweather station, through the transformation T : [0, 1] [0, 1]</p><p>T (Gh(x)) = Fh(x) (1)</p><p>Replacing x by G1h (u) in Equation (1) with u [0, 1]allows the following definition for the transform T :</p><p>T (u) = Fh(G1h (u)) (2)</p><p>Assuming that this later relationship remains valid in thefuture (i.e. Ff = T (Gf )), the researched CDF is givenby:</p><p>Ff (x) = Fh(G1h (Gf (x))) (3)</p><p>Following Michelangeli et al. (2009), the CDF-t is thendefined in two steps. First, estimates of (Fh, G1h , Gf )are non-parametrically modelled. Then, their combinationusing Equation (3) provides an estimate of Ff . Unlike theclassical quantile-matching approach which projects thesimulated future large-scale values on the historical CDFto compute and match quantiles, CDF-t takes into accountthe change in the large-scale CDF from the historical tothe future period. To downscale rainfall/surface tempera-tures over the full annual cycle, CDF-t is applied at eachgrid point to multiannual chronicles consisting of dailyrainfall/surface temperatures for each month of the cal-endar year (all January months, February months, etc.).The resulting local-scale daily chronicles of each month</p><p>are then rearranged to produce yearly estimates for theperiod chosen.</p><p>3. Calibration and validation over a historicalperiod</p><p>The CDF-t method is first calibrated and validated over aso-called historical period (19711999) which is cut intotwo contiguous temporal windows, 19711985 used forcalibration and 19861999 for which downscaled resultsare evaluated regarding local observations from IMD.The resulting downscaled CDFs are evaluated at eachgrid point against IMD data using KolmogorovSmirnov(KS) statistics providing an estimate of the maximumdifference between downscaled and observed local CDFs(Darling, 1957). The results for both rainfall and surfacetemperatures over southern India are presented. However,because rainfall has a discontinuity in zero, KS scoreshave also been computed after removal of days withno precipitation, this analysis is further discussed inSection 3.1. In addition, dedicated diagnostics are donefor three watersheds ranging from arid (Pandam Eru),semi-arid (Kudaliar) to wetter conditions (South Gundal).These locations are chosen to illustrate and discussthe performance of the method for different climaticenvironments.</p><p>3.1. Precipitation regimesFigure 1 presents KS statistics between CDFs fromoriginal GCMs rainfall (bi-linearly interpolated on the0.5 IMD grid) as well as from downscaled GCMs dataand IMD observed precipitation for 19861999 withcalibration of the CDF-t over the 19711985 period.As mentioned previously, this KS test is performed foreach grid point; therefore, the box-plots displayed inFigure 1 represent the spatial dispersion of the KS scoresfor the whole of the South India domain. For bothperiods, the different GCMs downscaled rainfall dataare characterized by a spatial dispersion comparable tooriginal GCM outputs. For the full year (top panels)and for all GCMs, even if they stay above the levelof statistical significance (0.019 at 0.05 significancelevel, not plotted), KS scores are substantially improvedfor downscaled daily rainfall compared to raw GCMsdata. Similar results are found for the monsoon period.Maximum KS scores computed in Figure 1 actually occurat the discontinuity of rainfall in zero, and the aboveresults thus show that dry days are better represented indownscaled fields than in original GCM outputs whencompared to observation. This is consistent with thefact that dry days are generally rare in GCMs data.Consequently, similar KS scores have been plotted inFigure 2 after removal of days with no rainfall (zerovalues) in all precipitation dataset. For the...</p></li></ul>

Recommended

View more >